If you use MCP servers to connect to multiple tools, you’re probably wasting most of your context window.
Each tool loads its input/output schemas into context. Just 5 tools – like Notion, Gmail, WhatsApp, Jira, and Confluence – can easily fill the entire 100k token window
And your AI agent gets dumber, slower, just because they can't hold large context at once.
Moreover, if you use OpenRouter or any API you will have to pay for storing all of this data in the context, meaning by default any user generates 100k+ context per message
How I fixed that:
- get-tools loads only the tools you need, using vector search
- Smaller context → up to 10x faster request
Meaning I can now connect 100+ different tools, and use them dynamically, without bloating the context window
alex_varga|10 months ago
Each tool loads its input/output schemas into context. Just 5 tools – like Notion, Gmail, WhatsApp, Jira, and Confluence – can easily fill the entire 100k token window
And your AI agent gets dumber, slower, just because they can't hold large context at once.
Moreover, if you use OpenRouter or any API you will have to pay for storing all of this data in the context, meaning by default any user generates 100k+ context per message How I fixed that: - get-tools loads only the tools you need, using vector search - Smaller context → up to 10x faster request Meaning I can now connect 100+ different tools, and use them dynamically, without bloating the context window